05 · Offer
Private AI Lab
Private models, measured — plus the data and evals to make them yours. When data must not leave your infrastructure, the model has to come to the data. We select, quantise and serve open models on your hardware, build the datasets and evaluation sets that prove they work — and work upstream in the open-source projects our own factory runs on.
The problem
Public-sector, health and financial data is often not allowed to go to a cloud model. And an open model that looks good on a leaderboard can still fail at your task. You need a model that fits your hardware, a measured serving profile, and test data from your own domain that shows whether the model is good enough — before an agent relies on it.
What we deliver
- Model selection for your task, data rules and hardware — tested on your cases, not on leaderboards
- Optimisation and quantisation of open models, with a measured serving profile: memory, speed, quality
- Private deployment in your infrastructure, with agents and access controls on top
- Datasets & evaluation: domain and synthetic training data, curation and quality filtering
- Evaluation sets and regression gates that every model and every update must pass
- LoRA fine-tuning of open models — where your evaluation shows it beats a well-prompted base model
- Open-source work upstream: issues reported and reproduced, features proposed in the projects we depend on
How it runs
Audit → Build → Retainer: the same line for every offer, tailored to this one. You decide after each phase whether to continue.
- 01
Audit
Data rules, hardware, use cases and a first evaluation set: which models are candidates, and what “good enough” means for your task.
- 02
Build
Model build and serving profile, private deployment, datasets and evaluation sets — and LoRA fine-tuning where the evaluation shows it pays off.
- 03
Retainer
New open models tested against your gate as they appear, updates rolled out as releases, evaluation sets grown with real cases.
Matching industries
Where this offer meets ground we know from the inside.
Questions
Can our data stay in-house?
Yes. We build on local and self-hosted models when data must not leave your infrastructure, and we scope every tool an agent can reach.
Do you fine-tune models?
We offer LoRA fine-tuning of open models as a service, when an evaluation shows it beats a well-prompted base model. Our public LoRA adapter for Qwen3-4B is a fine-tune on our own synthetic dataset; our other public models are quantised builds of existing open models, not fine-tunes — and we always say which is which.
What hardware do we need?
That depends on the model and the load — which is why every build comes with a measured serving profile. One example from our published builds: a 125-billion-parameter mixture-of-experts model that runs on a single Mac with about 40 GB of memory.
Why build our own evaluation set?
Because public benchmarks do not know your documents, your rules or your language. A few dozen real cases with known-correct answers show quickly whether a model is good enough — and they become the regression gate for every later update.
Talk it through in 30 minutes.
Bring one process and your stack. We come prepared and tell you honestly whether an agent pays off — and which offer is the right start.
Or write to us: office@novaeon.studio